AHAL AI
No LLM judges an LLM.
- AI / Developer Tools
- Designer
- Design and whitepaper
Context
AHAL AI is an engineering risk intelligence platform. It predicts the blast radius of a pull request, diagnoses production root causes, and stages autonomous remediation.
Approach
Every LLM-proposed change is gated behind deterministic Tree-sitter and AST verification. There is no LLM-judges-LLM step anywhere in the loop. This is the deterministic core and LLM language layer split applied to code changes.
System
Multi-agent repository ingestion
Triggered by GitHub webhooks
LLM-proposed change
Remediation is proposed, never trusted
Tree-sitter / AST verification
Deterministic gate on every proposed change
Staged autonomous remediation
Only verified changes proceed
Build
- PR blast-radius prediction
- Production root-cause diagnosis
- Staged autonomous remediation
- Multi-agent repository ingestion with GitHub webhook automation
- Technical whitepaper, benchmarked against a competing multi-agent code-review system
Challenges
- Letting an LLM propose changes without letting it approve them: every proposal goes through deterministic Tree-sitter and AST verification.
- Ingesting a whole repository across multiple agents, then reacting to GitHub webhooks as changes arrive.
- Showing the design holds up by benchmarking it against a competing multi-agent code-review system.
Result
Designed the platform and authored the technical whitepaper, benchmarking the design against a competing multi-agent code-review system.
Learnings
A verifier that shares the generator's failure modes is not a verifier. Determinism is what makes the gate worth having.